Learning Deep Models from Weak Labels for Water Surface Segmentation in SAR Images

"Learning Deep Models from Weak Labels for Water Surface Segmentation in SAR Images" paper data: IGARSS21_samples.npy: a list of 100 input weak-label pairs (as tuples) for testing the architecture presented in the paper. The input is a 2-by-128-by-128 tensor cast to np.float32 dtype, where the two channels represent the VH and VV polarization. Each tensor is sampled from S1-IW \(\LARGE{\sigma_{0}}\) intensity images (resampled on a 10-by-10 meters grid) acquired over the validation area pointed out in the paper. The \(\LARGE{\sigma_{0}}\) intensity values are clipped to 1 and scaled by a constant factor of 100. The weak label is a 128-by-128 matrix cast to np.uint8 dtype, whose pixels can have only one of the two following values: 1=water body, 0=land. The source of the weak labels is the aggregated (see paper) 2015 Copernicus’ Water and Wetness High Resolution Layer. IGARSS21_weights.pth: weights of the best architecture presented in the paper.